ActEval

ActEval is an insurance-oriented Python toolkit for evaluating caller-supplied prediction arrays. It keeps accuracy, calibration, discrimination, tail, uncertainty, and realized-consequence objectives separate instead of collapsing them into a universal score.

Install

python -m pip install acteval-insurance

Add plotting or numerical Tweedie distribution support only when needed:

python -m pip install "acteval-insurance[plot,tweedie]"

First evaluation

import acteval as ae

result = ae.evaluate(
    y_true=[0.0, 0.4, 1.0, 2.0, 4.0],
    y_pred=[0.1, 0.5, 0.9, 1.8, 3.6],
    exposure=[1.0, 0.5, 1.2, 0.8, 1.5],
    input_scale="rate",
    task="claim_frequency",
    context={"model_id": "frequency-glm-v4", "split": "holdout"},
)

print(result.to_dataframe())

Start with the API guide for input and result contracts, then use the metric reference to choose diagnostics that match the prediction functional and portfolio question.

What this documentation covers

• Point and predictive-distribution evaluation
• Explicit exposure-scale handling
• Metric-specific model comparison
• Bootstrap sampling uncertainty
• Segment, temporal, and prediction-drift reports
• Immutable result metadata and exports
• Illustrative realized-consequence helpers

Scope

ActEval is not a complete model-validation or governance system. It cannot detect data leakage, validate a train/test split, choose portfolio thresholds, or replace feature-aware conditional calibration. The compatibility policy describes software guarantees, not actuarial fitness or production readiness.